Artificial intelligence for direct-to-physician reporting of ambulatory electrocardiography
Linda Johnson, Piotr Zadrozniak, Grzegorz Jasina, Agnieszka Grotek-Cuprjak, Jason G. Andrade, Emma Svennberg, Søren Zöga Diederichsen, William F. McIntyre, Stavros Stavrakis, Juan Benezet‐Mazuecos, Philipp Krisai, Zaza Iakobishvili, A. Laish-Farkash, Sanjeev P. Bhavnani, Erik Ljungström, Justinas Bacevičius, Nick L. van Vreeswijk, Michiel Rienstra, Raphael Spittler, J. A. Marx, Alireza Oraii, Ángel Miracle Blanco, Ada Sánchez Lozano, Irina Mustafina, Stefanos Zafeiropoulos, Richard G. Bennett, Jasmine Bisson, Dominik Linz, Yonatan Kogan, Evan S. Glazer, Gergana Marincheva, Michael Rahkovich, Einat Shaked, Martin H. Ruwald, Ketil Haugan, Jakub Weclawski, Glauco Radoslovich, Shahin Jamal, Axel Brandes, Paweł T. Matusik, Martin Manninger, Pascal Meyre, Steffen Blum, Anders Persson, Alexandra Måneheim, Per Hammarlund, Artur Fedorowski, Tigist Wodaje, Christian Lewinter, Vytautas Juknevičius, Rusne Jakaite, Christine Shen, Taya V. Glotzer, Pyotr G. Platonov, Gunnar Engström, Alexander P. Benz, Jeff S. Healey
Abstract
Developments in ambulatory electrocardiogram (ECG) technology have led to vast amounts of ECG data that currently need to be interpreted by human technicians. Here we tested an artificial intelligence (AI) algorithm for direct-to-physician reporting of ambulatory ECGs. Beat-by-beat annotation of 14,606 individual ambulatory ECG recordings (mean duration = 14 ± 10 days) was performed by certified ECG technicians (n = 167) and an ensemble AI model, called DeepRhythmAI. To compare the performance of the AI model and the technicians, a random sample of 5,235 rhythm events identified by the AI model or by technicians, of which 2,236 events were identified as critical arrhythmias, was selected for annotation by one of 17 cardiologist consensus panels. The mean sensitivity of the AI model for the identification of critical arrhythmias was 98.6% (95% confidence interval (CI) = 97.7-99.4), as compared to 80.3% (95% CI = 77.3-83.3%) for the technicians. False-negative findings were observed in 3.2/1,000 patients for the AI model versus 44.3/1,000 patients for the technicians. Accordingly, the relative risk of a missed diagnosis was 14.1 (95% CI = 10.4-19.0) times higher for the technicians. However, a higher false-positive event rate was observed for the AI model (12 (interquartile range (IQR) = 6-74)/1,000 patient days) as compared to the technicians (5 (IQR = 2-153)/1,000 patient days). We conclude that the DeepRhythmAI model has excellent negative predictive value for critical arrhythmias, substantially reducing false-negative findings, but at a modest cost of increased false-positive findings. AI-only analysis to facilitate direct-to-physician reporting could potentially reduce costs and improve access to care and outcomes in patients who need ambulatory ECG monitoring.